Table of Contents

1  Ground predictions
1.1  PVLib Clearsky
2  Train/test on NSRDB data to find optimal parameters
3  Train on all NSRDB data, test various freq of ground data
3.1  30 min freq ground data
3.2  15 min freq ground data
3.3  10 min freq ground data
3.4  5 min freq ground data
3.5  1 min freq ground data
4  Save model
5  Conclusion

In [1]:
import pandas as pd
import numpy as np
import os
import datetime
import matplotlib
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1 import make_axes_locatable
from sklearn import tree
from sklearn import ensemble

import pytz
import itertools
import visualize
import utils
import pydotplus
import xgboost as xgb

from sklearn import metrics
from sklearn import model_selection

import pvlib
import cs_detection

import visualize_plotly as visualize
from plotly.offline import download_plotlyjs, init_notebook_mode, plot, iplot
import plotly.graph_objs as go
init_notebook_mode(connected=True)

from IPython.display import Image

%load_ext autoreload
%autoreload 2

np.set_printoptions(precision=4)
%matplotlib notebook

Ground predictions

PVLib Clearsky

Only making ground predictions using PVLib clearsky model and statistical model. NSRDB model won't be available to ground measurements.

In [2]:
nsrdb = cs_detection.ClearskyDetection.read_pickle('ornl_nsrdb_1.pkl.gz')
nsrdb.df.index = nsrdb.df.index.tz_convert('EST')
nsrdb.trim_dates('01-01-2014', '01-01-2016')
In [3]:
nsrdb.time_from_solar_noon('Clearsky GHI pvlib', 'tfn')
In [4]:
ground = cs_detection.ClearskyDetection.read_pickle('ornl_ground_1.pkl.gz')
ground.df.index = ground.df.index.tz_convert('EST')
In [5]:
feature_cols = [
'tfn',
'abs_ideal_ratio_diff',
'abs_ideal_ratio_diff mean',
'abs_ideal_ratio_diff std',
'abs_ideal_ratio_diff max',
'abs_ideal_ratio_diff min',
'GHI Clearsky GHI pvlib gradient ratio', 
'GHI Clearsky GHI pvlib gradient ratio mean', 
'GHI Clearsky GHI pvlib gradient ratio std', 
'GHI Clearsky GHI pvlib gradient ratio min', 
'GHI Clearsky GHI pvlib gradient ratio max', 
'GHI Clearsky GHI pvlib gradient second ratio', 
'GHI Clearsky GHI pvlib gradient second ratio mean', 
'GHI Clearsky GHI pvlib gradient second ratio std', 
'GHI Clearsky GHI pvlib gradient second ratio min', 
'GHI Clearsky GHI pvlib gradient second ratio max', 
'GHI Clearsky GHI pvlib line length ratio',
'GHI Clearsky GHI pvlib line length ratio gradient',
'GHI Clearsky GHI pvlib line length ratio gradient second'
]

target_cols = ['sky_status']

Train/test on NSRDB data to find optimal parameters

In [66]:
train = cs_detection.ClearskyDetection(nsrdb.df)
train.trim_dates('01-01-2010', '01-01-2015')
test = cs_detection.ClearskyDetection(nsrdb.df)
test.trim_dates('01-01-2015', None)
In [67]:
train.scale_model('GHI', 'Clearsky GHI pvlib', 'sky_status')
In [68]:
utils.calc_all_window_metrics(train.df, 3, meas_col='GHI', model_col='Clearsky GHI pvlib', overwrite=True)
In [60]:
param_grid = {'max_depth': [3, 6, 9, 12], 'n_estimators': [64, 128, 256]}
import itertools import warnings best_score = 0 best_params = {} with warnings.catch_warnings(): warnings.simplefilter('ignore') for depth, n_est in itertools.product(param_grid['max_depth'], param_grid['n_estimators']): # clf = xgb.XGBClassifier(max_depth=depth, n_estimators=n_est, learning_rate=lr, n_jobs=4) clf = ensemble.RandomForestClassifier(max_depth=depth, n_estimators=n_est, n_jobs=-1) clf.fit(train.df[feature_cols].values, train.df[target_cols].values.flatten()) test = cs_detection.ClearskyDetection(nsrdb.df) test.trim_dates('06-01-2015', None) pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 3) score = metrics.f1_score(test.df['sky_status'], pred) indicator = '' if score > best_score: best_score = score best_params['max_depth'] = depth best_params['n_estimators'] = n_est indicator = '*' print('max_depth: {}, n_estimators: {}, accuracy: {} {}'.format(depth, n_est, score, indicator))max_depth: 5, n_estimators: 300, learning_rate: 0.1, accuracy: 0.910144927536232 *
In [61]:
best_params = {'max_depth': 6, 'n_estimators': 128}
In [62]:
# clf = xgb.XGBClassifier(**best_params, n_jobs=4)
clf = ensemble.RandomForestClassifier(**best_params, n_jobs=-1)
clf.fit(train.df[feature_cols].values, train.df[target_cols].values.flatten())
Out[62]:
RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
            max_depth=6, max_features='auto', max_leaf_nodes=None,
            min_impurity_split=1e-07, min_samples_leaf=1,
            min_samples_split=2, min_weight_fraction_leaf=0.0,
            n_estimators=128, n_jobs=-1, oob_score=False,
            random_state=None, verbose=0, warm_start=False)
In [88]:
test = cs_detection.ClearskyDetection(nsrdb.df)
test.trim_dates('01-01-2015', None)
In [89]:
%%time
pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 3, multiproc=True, by_day=True).astype(bool)

CPU times: user 4.35 s, sys: 656 ms, total: 5 s
Wall time: 42.6 s
In [90]:
test = cs_detection.ClearskyDetection(nsrdb.df)
test.trim_dates('01-01-2015', None)
In [91]:
%%time
pred2 = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 3, multiproc=False, by_day=True).astype(bool)

/Users/benellis/duramat/clearsky_detection/utils.py:229: RuntimeWarning:

Large scaling value.  Day will not be further assessed or scaled.

/Users/benellis/duramat/clearsky_detection/utils.py:236: RuntimeWarning:

Scaling did not converge.


CPU times: user 1min 4s, sys: 3.28 s, total: 1min 7s
Wall time: 2min 26s
In [14]:
vis = visualize.Visualizer()
vis.add_line_ser(test.df['GHI'], 'GHI')
vis.add_line_ser(test.df['Clearsky GHI pvlib'], 'GHI_cs')
vis.add_circle_ser(test.df[(test.df['sky_status'] == 0) & (pred)]['GHI'], 'ML clear only')
vis.add_circle_ser(test.df[(test.df['sky_status'] == 1) & (~pred)]['GHI'], 'NSRDB clear only')
vis.add_circle_ser(test.df[(test.df['sky_status'] == 1) & (pred)]['GHI'], 'ML+NSRDB clear only')
vis.show()
In [92]:
pred.equals(pred2)
Out[92]:
True
In [93]:
cm = metrics.confusion_matrix(test.df['sky_status'].values, pred)
vis = visualize.Visualizer()
vis.plot_confusion_matrix(cm, labels=['cloudy', 'clear'])
In [94]:
print(metrics.f1_score(test.df['sky_status'].values, pred))

0.891558935361

Train on all NSRDB data, test various freq of ground data

In [95]:
train = cs_detection.ClearskyDetection(nsrdb.df)
train.scale_model('GHI', 'Clearsky GHI pvlib', 'sky_status')
utils.calc_all_window_metrics(train.df, 3, meas_col='GHI', model_col='Clearsky GHI pvlib', overwrite=True)
clf.fit(train.df[feature_cols].values, train.df[target_cols].values.flatten())
Out[95]:
RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
            max_depth=6, max_features='auto', max_leaf_nodes=None,
            min_impurity_split=1e-07, min_samples_leaf=1,
            min_samples_split=2, min_weight_fraction_leaf=0.0,
            n_estimators=128, n_jobs=-1, oob_score=False,
            random_state=None, verbose=0, warm_start=False)
In [96]:
bar = go.Bar(x=feature_cols, y=clf.feature_importances_)
iplot([bar])

30 min freq ground data

In [97]:
ground = cs_detection.ClearskyDetection.read_pickle('ornl_ground_1.pkl.gz')
ground.df.index = ground.df.index.tz_convert('EST')
test = cs_detection.ClearskyDetection(ground.df)
In [98]:
test.trim_dates('10-01-2008', '10-16-2008')
In [99]:
test.time_from_solar_noon('Clearsky GHI pvlib', 'tfn')
In [100]:
test.df = test.df[test.df.index.minute % 30 == 0]
In [101]:
pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 3, multiproc=True, by_day=True).astype(bool)
In [102]:
train2 = cs_detection.ClearskyDetection(train.df)
train2.intersection(test.df.index)
In [103]:
nsrdb_clear = train2.df['sky_status'].values
ml_clear = pred
vis = visualize.Visualizer()
vis.add_line_ser(test.df['GHI'], 'GHI')
vis.add_line_ser(test.df['Clearsky GHI pvlib'], 'GHI_cs')
vis.add_circle_ser(test.df[ml_clear & ~nsrdb_clear]['GHI'], 'ML clear only')
vis.add_circle_ser(test.df[~ml_clear & nsrdb_clear]['GHI'], 'NSRDB clear only')
vis.add_circle_ser(test.df[ml_clear & nsrdb_clear]['GHI'], 'Both clear')
vis.show()

15 min freq ground data

In [111]:
ground = cs_detection.ClearskyDetection.read_pickle('ornl_ground_1.pkl.gz')
ground.df.index = ground.df.index.tz_convert('EST')
test = cs_detection.ClearskyDetection(ground.df)
In [112]:
test.trim_dates('10-01-2008', '10-16-2008')
In [113]:
test.time_from_solar_noon('Clearsky GHI pvlib', 'tfn')
In [114]:
test.df = test.df[test.df.index.minute % 15 == 0]
In [115]:
pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 5, multiproc=True, by_day=True).astype(bool)
In [116]:
train2 = cs_detection.ClearskyDetection(train.df)
train2.trim_dates('10-01-2008', '10-16-2008')
train2.df = train2.df.reindex(pd.date_range(start=train2.df.index[0], end=train2.df.index[-1], freq='15min'))
train2.df['sky_status'] = train2.df['sky_status'].fillna(False)
In [117]:
nsrdb_clear = train2.df['sky_status']
ml_clear = test.df['sky_status iter']
vis = visualize.Visualizer()
vis.add_line_ser(test.df['GHI'], 'GHI')
vis.add_line_ser(test.df['Clearsky GHI pvlib'], 'GHI_cs')
vis.add_circle_ser(test.df[ml_clear & ~nsrdb_clear]['GHI'], 'ML clear only')
vis.add_circle_ser(test.df[~ml_clear & nsrdb_clear]['GHI'], 'NSRDB clear only')
vis.add_circle_ser(test.df[ml_clear & nsrdb_clear]['GHI'], 'Both clear')
vis.show()

10 min freq ground data

In [33]:
ground = cs_detection.ClearskyDetection.read_pickle('ornl_ground_1.pkl.gz')
ground.df.index = ground.df.index.tz_convert('EST')
test = cs_detection.ClearskyDetection(ground.df)
In [34]:
test.trim_dates('10-01-2008', '10-16-2008')
In [35]:
test.time_from_solar_noon('Clearsky GHI pvlib', 'tfn')
In [36]:
test.df = test.df[test.df.index.minute % 10 == 0]
In [37]:
pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 7, multiproc=True, by_day=True).astype(bool)

/Users/benellis/duramat/clearsky_detection/utils.py:229: RuntimeWarning:

Large scaling value.  Day will not be further assessed or scaled.

/Users/benellis/duramat/clearsky_detection/utils.py:236: RuntimeWarning:

Scaling did not converge.
In [38]:
train2 = cs_detection.ClearskyDetection(train.df)
train2.trim_dates('10-01-2008', '10-16-2008')
train2.df = train2.df.reindex(pd.date_range(start=train2.df.index[0], end=train2.df.index[-1], freq='10min'))
train2.df['sky_status'] = train2.df['sky_status'].fillna(False)
In [39]:
nsrdb_clear = train2.df['sky_status']
ml_clear = test.df['sky_status iter']
vis = visualize.Visualizer()
vis.add_line_ser(test.df['GHI'], 'GHI')
vis.add_line_ser(test.df['Clearsky GHI pvlib'], 'GHI_cs')
vis.add_circle_ser(test.df[ml_clear & ~nsrdb_clear]['GHI'], 'ML clear only')
vis.add_circle_ser(test.df[~ml_clear & nsrdb_clear]['GHI'], 'NSRDB clear only')
vis.add_circle_ser(test.df[ml_clear & nsrdb_clear]['GHI'], 'Both clear')
vis.show()

5 min freq ground data

In [118]:
ground = cs_detection.ClearskyDetection.read_pickle('ornl_ground_1.pkl.gz')
ground.df.index = ground.df.index.tz_convert('EST')
test = cs_detection.ClearskyDetection(ground.df)
In [119]:
test.trim_dates('10-01-2008', '10-16-2008')
In [120]:
test.time_from_solar_noon('Clearsky GHI pvlib', 'tfn')
In [121]:
test.df = test.df[test.df.index.minute % 5 == 0]
In [122]:
pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 13, multiproc=True, by_day=True).astype(bool)
In [123]:
train2 = cs_detection.ClearskyDetection(train.df)
train2.trim_dates('10-01-2008', '10-16-2008')
train2.df = train2.df.reindex(pd.date_range(start=train2.df.index[0], end=train2.df.index[-1], freq='5min'))
train2.df['sky_status'] = train2.df['sky_status'].fillna(False)
In [124]:
nsrdb_clear = train2.df['sky_status']
ml_clear = test.df['sky_status iter']
vis = visualize.Visualizer()
vis.add_line_ser(test.df['GHI'], 'GHI')
vis.add_line_ser(test.df['Clearsky GHI pvlib'], 'GHI_cs')
vis.add_circle_ser(test.df[ml_clear & ~nsrdb_clear]['GHI'], 'ML clear only')
vis.add_circle_ser(test.df[~ml_clear & nsrdb_clear]['GHI'], 'NSRDB clear only')
vis.add_circle_ser(test.df[ml_clear & nsrdb_clear]['GHI'], 'Both clear')
vis.add_line_ser(test.df['abs_ideal_ratio_diff'] * 100)
vis.show()

1 min freq ground data

In [125]:
ground = cs_detection.ClearskyDetection.read_pickle('ornl_ground_1.pkl.gz')
ground.df.index = ground.df.index.tz_convert('EST')
test = cs_detection.ClearskyDetection(ground.df)
In [126]:
test.trim_dates('10-01-2008', '10-16-2008')
In [127]:
test.time_from_solar_noon('Clearsky GHI pvlib', 'tfn')
In [128]:
test.df = test.df[test.df.index.minute % 1 == 0]
In [129]:
pred = test.iter_predict_daily(feature_cols, 'GHI', 'Clearsky GHI pvlib', clf, 61, multiproc=True, by_day=True).astype(bool)
In [130]:
train2 = cs_detection.ClearskyDetection(train.df)
train2.trim_dates('10-01-2008', '10-16-2008')
train2.df = train2.df.reindex(pd.date_range(start=train2.df.index[0], end=train2.df.index[-1], freq='1min'))
train2.df['sky_status'] = train2.df['sky_status'].fillna(False)
In [131]:
nsrdb_clear = train2.df['sky_status']
ml_clear = test.df['sky_status iter']
vis = visualize.Visualizer()
vis.add_line_ser(test.df['GHI'], 'GHI')
vis.add_line_ser(test.df['Clearsky GHI pvlib'], 'GHI_cs')
vis.add_circle_ser(test.df[ml_clear & ~nsrdb_clear]['GHI'], 'ML clear only')
vis.add_circle_ser(test.df[~ml_clear & nsrdb_clear]['GHI'], 'NSRDB clear only')
vis.add_circle_ser(test.df[ml_clear & nsrdb_clear]['GHI'], 'Both clear')
vis.add_line_ser(test.df['abs_ideal_ratio_diff'] * 100)
vis.show()

Save model

In [54]:
import pickle
In [55]:
with open('ornl_trained.pkl', 'wb') as f:
    pickle.dump(clf, f)
In [56]:
!ls ornl*

ornl_ground.pkl      ornl_nsrdb.pkl       ornl_trained.pkl
ornl_ground.pkl.gz   ornl_nsrdb.pkl.gz
ornl_ground_1.pkl.gz ornl_nsrdb_1.pkl.gz

Conclusion

In general, the clear sky identification looks good. At lower frequencies (30 min, 15 min) we see good agreement with NSRDB labeled points. I suspect this could be further improved my doing a larger hyperparameter search, or even doing some feature extraction/reduction/additions.

In [ ]: